Deep learning based CT images automatic analysis model for active/non-active pulmonary tuberculosis differential...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonology

Background:

• Active pulmonary tuberculosis (ATB) is highly infectious and fatal, necessitating rapid diagnosis.
• Current diagnostic methods (bacteriology, culturing, radiology) are slow and labor-intensive.
• AI offers a potential solution for faster and more accurate ATB diagnosis.

Purpose of the Study:

• To develop and validate an AI model for the differential diagnosis of active pulmonary tuberculosis (ATB) using CT scans.
• To compare the performance of different deep learning models for ATB diagnosis.

Main Methods:

• Collected CT scans and clinical data from 1160 ATB and 1131 non-ATB patients.
• Utilized a 3D Nested UNet for lung field segmentation and 3D VGG-16, EfficientNet, and ResNet-50 for classification.
• Validated the model on internal and external test sets, including 100 ATB and 100 non-ATB cases in the external set.

Main Results:

• The 3D ResNet-50 model achieved an AUC of 0.961 (internal) and 0.946 (external).
• The AI model demonstrated higher diagnostic accuracy than experienced radiologists.
• The AI model diagnosed CT scans 10 times faster than human experts and visualized key lung lesions.

Conclusions:

• Developed an AI tool to assist in differentiating active pulmonary tuberculosis from non-ATB.
• The 3D ResNet-50 model shows significant potential for rapid and accurate bedside diagnosis.
• This AI tool can improve tuberculosis control by enabling timely diagnosis and treatment.